Slashdot Mirror
AI Can Detect Sexual Orientation Based On Person's Photo (cnbc.com)
ugen shares a report from CNBC: Artificial Intelligence (AI) can now accurately identify a person's sexual orientation by analyzing photos of their face, according to new research. The Stanford University study, which is set to be published in the Journal of Personality and Social Psychology and was first reported in The Economist, found that machines had a far superior "gaydar" when compared to humans. Slashdot reader randomlygeneratename adds: Researchers built classifiers trained on photos from dating websites to predict the sexual orientation of users. The best classifier used logistic regression over features extracted from a VGG-Face conv-net. The latter was done to prevent overfitting to background, non-facial information. Classical facial feature extraction also worked with a slight drop in accuracy. From multiple photos, they achieved an accuracy of 91% for men and 83% for women (and 81% / 71% for a single photo). Humans were only able to get 61% and 54%, respectively. One caveat is the paper mentions it only used Caucasian faces. The paper went on to discuss how this capability can be an invasion of privacy, and conjectured that other types of personal information might be detectable from photos. The source paper can be found here.
...definitely applies to this situation. This has some pretty negative implications in particularly homophobic regions. All the more reason not to visit the pacific northwestern US or the middle eastern region in general if this thing gets to be widespread.
Furries make the internet go.
If the AI were to simply assign scores of "Straight" to EVERYONE, it would achieve 90% accuracy for men and 85% accuracy for women, since about 90% of men are straight and about 85% of women are straight. So scores of 91% and 85% accuracy are not statistically significant.
First off, what they claim to have created is tantamount to computer-assisted phrenology -- long since debunked and tossed on the scrapheap of superstition.
The most obvious flaw appears here, starting on line 208:
*headdesk*
Leaving aside the gigantic issues presented by self-reporting and self-selecting samples, these idiots failed to account for a common practice among hetero women on dating sites, which is to falsely claim to be seeking other women as a means to reduce or eliminate an onslaught of tacky propositions from clueless het-boys.
Other glaring flaws include:
An actual sociologist could probably identify dozens of other flaws, any one of which would be fatal to the work.
I would undertake to create a similar piece of software that tries to identify criminals from photographs, and use police mugshots to train it. Surprise! Black people are more likely to be criminals! GIGO.
Frankly, I think they should have taken their theme from the closing paragraphs of their paper: "We created a digital phrenologist out of deep neural networks and other off-the-shelf parts that coughs up results that seem relevant and meaningful to the layman, when in fact they're utter garbage." That would have been a good paper.
Perhaps we can indeed learn new things by letting a DNN stare at human faces. But IMHO this paper is utterly valueless in identifying what those might be. GIGO.
Editor, A1-AAA AmeriCaptions
It was only 60 years ago that being homosexual was outlawed in most "civilized western nations" and "treated" with drugs and electroshock. Society doesn't have a direction of continued permissiveness and can slip back at any time. Imagine what this ability would then be used for.